Google claims its new AI agent AlphaEvolve is able to improve itself

Introduction

Google recently wrote a blog post about it's new "AI agent" AlphaEvolve, powered by Google's Gemini AI. It's an AI agent which evolves algorithms for math and practical applications in computing by combining the creativity of large language models with automated evaluators.

How AlphaEvolve works

AlphaEvolve uses Gemini's large language language models with an evolutionary loop and scoring system. It uses a combination of neural networks evolutionary algorithms. Here's how it operates:

1. User input: A user provides a program and evaluation function.
2. Mutation: Variations of the program are generated (A team of LLM's generate a set of different solutions)
3. Evaluation: Each variation is scored.
4. Evolution: The best variations are selected and evolved using Gemini models.

This process helps AlphaEvolve refine algorithms over generations.

Continuous self-improvement

AlphaEvolve learns and improves with each iteration. By evaluating and selecting effective algorithm variations, it adapts and evolves, leading to optimized solutions. This ensures AlphaEvolve stays advanced in algorithm design.

Real-world applications

AlphaEvolve is used in various areas within Google:

• Data center optimization: It developed a better scheduling heuristic for Borg, reclaiming 0.7% of resources.
• Hardware design: Proposed improvements for TPU circuits, speeding up matrix multiplication cores by 23% and reducing GPU execution times by 32.5%. Therefore optimizing the very alghoritms that power AlphaEvolve.
• Mathematical advancements: Improved 14 matrix multiplication benchmarks and made progress on over 20% of more than 50 open problems.

How AlphaEvolve compares with AlphaTensor

Unlike AlphaTensor, which focused on matrix multiplication, AlphaEvolve is a generalist. It can work on any problem with a machine-evaluable objective, such as mathematical limits or code optimization.

AlphaEvolve is a great example of how LLM's can be used to not only create great solutions but also improve existing alghorithms and solutions, which up to know hasn't been done before with LLM's.